The International Seed Testing Association has as one of its important objectives to develop and test methods used to quantify seed germination. Existing methods to test seed viability include X-ray analysis, tetrazolium staining, a cut test, in some cases the careful extraction and sterile culture of zygotic embryos under aseptic in vitro conditions, but germination tests are probably the most widely used. Each method has advantages and disadvantages and consequently no single method provides a definitive means of accurately quantifying seed viability. For example, while relatively new and non destructive the X-ray assessment of seeds only indicates whether seeds are structurally intact , not whether the seeds are actually alive so while the seeds can be placed back into storage following assessment the use of this test on its own does not provide any certainty as to whether any structurally sound seeds are specifically alive. Although the germination test is probably one of the most accurate ways to assess seed lots, it is labor intensive, time consuming , and it may dependent upon knowledge concerning the optimal germination conditions and dormancy breaking requirements of the species in question. In addition, a significant concern from a seed bank management perspective is the fact that germination tests are ‘‘irreversible’’, so if seeds are viable and therefore germinate, the genetic information that they represent will be lost if not grown to maturity which may not possible in large seed banks. This is a particular concern with critically endangered species, as seed stocks may be exceptionally rare and in some cases impossible to source again. With currently available seed viability tests being largely irreversible, generally destructive, and sensitive to subjective data interpretation,hydroponic grow systems there is a pressing need for development of non-destructive and quantitative methods to assess viability and germination of precious seed stocks.
Several studies have demonstrated the potential of reflectance based spectroscopy methods in studies of plant seeds, including detection of internal infestations by weevils in dry field peas, classification of near isogenic maize lines, ageing of cabbage seeds, classification of near isogenic maize lines, differentiation between black walnut shell and pulp, sorting of lettuce seeds , and viability of horticultural seeds. These spectroscopy studies are based on the fundamental assumption that reflectance data acquired from the seed coat provides indicative information about the quality/ germination of the given seed. The research objective is therefore to identify portions of the wavelength spectrum, in which seeds show a significant and measurable change in certain parts of the examined reflectance spectrum and associated that change in reflectance with certain traits, such as, germination . A wide range of classification methods have been used as part of using reflectance data to characterize seeds and food products; these classification methods include support vector machine, variogram analysis, partial least square analysis, and linear discriminant analysis. LDA is based on discriminant functions, which are linear combinations of features and with one function for each target class. For each observation, a discriminant score is calculated and the observation is assigned to the class for which the discriminant function generates the highest discriminant score. In this study, we used hyperspectral imaging data to determine the germination of seeds from three native Australian tree species [Acacia cowleana Tate , Banksia prionotes L.F. , and Corymbia calophylla K.D. Hill & L.A.S. Johnson ]. These species represent the Australian flora with many different species held in seed banks across Australia as well as internationally including a number of critically endangered taxa. Seeds were exposed to standardized rapid ageing conditions, and at each assessment point were subjected to germination testing and hyperspectral imaging. We hypothesized that there would be detectable difference in seed coat reflectance between germinating and nongerminating seeds, and that changes in reflectance profiles would be most pronounced in spectral bands near known pigment peaks involved in photosynthesis and/or near spectral bands used in published indices to predict chlorophyll or nitrogen content in leaves.
The potential benefits of developing accurate machine vision systems to automate non-destructive monitoring of seed germination are discussed in the context of management of seed banks, botanic gardens, and implementation of vegetation restoration programs.A push-broom hyperspectral camera was mounted 40 cm above the seeds, and hyperspectral images were acquired with the spatial resolution of 50 pixels per mm2 under artificial lighting . The main specifications of the hyperspectral camera were: Firewire interface , 12 bit digital output, 240 spectral bands from 392 to 889 nm by 640 pixels . The objective lens had a 35 mm focal length with a 7 field of view, and it was optimized for the near-infrared and visible near-infrared spectra. During hyperspectral image acquisition in the lab, RH was between 30% and 40% and temperature 19–22C. A piece of white Teflon was used for white calibration. reflectance value was referred to relative reflectance and compared to that obtained from white Teflon. Colored plastic cards were imaged at all hyperspectral imaging events, and average reflectance profiles from these cards were used to confirm high consistency of hyperspectral image acquisition conditions .Immediately after hyperspectral imaging at each sampling time, the sub-samples of seeds were transferred to Petri dishes with semi solidified water agar for germination testing. Each seed was sequentially placed onto the dish and individually labelled with a permanent marker . Petri dishes were sealed with plastic film, wrapped in aluminium foil and incubated in a 15C growth chamber and checked weekly for germination. For determination of the predicted time required for germination to decline to 50% probit analysis was performed in Genstat version 10.0 . Seeds were aged at 60 C rather than 45C as is more common in seed ageing experiments. P50 values were calculated using the original data then adjusted to 45C by multiplying 60C P50 values by 8.44 as described in previous studies.Pixels with R660 reflectance values outside the stipulated ranges were excluded before average reflectance profiles were generated for each seed. We obtained 200 average reflectance profiles from each species , and relationships between results from the germination tests and hyperspectral imaging data were tested using LDA classification.
To select an optimized subset of the 10 ‘‘best’’ spectral bands, we conducted a forward stepwise LDA based on all average reflectance profiles for each of the three seeds species, and only these 10 spectral bands were used. Although not presented in this study, we conducted additional classification analyses with both more or fewer spectral bands,rolling benches and negligible classification accuracy was gained by including more than 10 spectral bands, and significant classification accuracy was lost by reducing the number of spectral bands. Thus, using the ‘‘best’’ 10 spectral bands was considered an optimum for this particular application. classification accuracies of each LDA was based on independent validation, as the original 200 average reflectance profiles and germination test results from each seed species were randomly divided into 80% training data and 20% independent validation data, This random division of data into training and validation data was repeated five times, and we calculated the average classification accuracy from the five randomized divisions of each data set.The effect of experimental ageing showed that within 0– 10 days, germination of Acacia and Corymbia seeds was above 90%, but both species of seeds showed a considerable decrease in germination after 10–20 or 20–30 days of experimental ageing, respectively . From the onset of ageing, there was an exponential decline in germination of Banksia seeds, and none of the seeds germinated after 30 days of experimental ageing. Germination for all three species was below 10% after 30 days of ageing, with B. prionotes seeds showing the steepest decline in germination with a P50 of only 7.0 days . For comparison, the P50 values for A. cowleana and C. calophylla seeds were 19.3 and 22.9 days respectively. Adjusted P50 values for all three species were 163.3, 59.1 and 193.3 days for C. calophylla, A. cowleana and B. prionotes respectively. With 25 seeds and eight sampling events , we obtained the following numbers of germinating and non-germinating seeds: Acacia , Banksia , and Corymbia .Average reflectance profiles acquired after 5 , 20 , and 20–30 days ofstandardized rapid ageing were used to illustrate the difference in reflectance between germinating and non-germinating seeds on days with similar numbers of seeds in both categories. Average reflectance profiles from the seed coat of the three species showed similar pattern with relative reflectance values commencing around 0.10 in spectral bands near 400 nm and gradually increasing reflectance in spectral bands near 900 nm. Loss of germination caused a decrease in reflectance in Banksia and Corymbia seeds, while it caused an increase in reflectance in Acacia seeds. We wish to highlight that it was virtually impossible to distinguish germinating and non-germinating seeds on the basis of visual inspection.
This statement is confirmed by the fact that reflectance values across the visual part of the spectrum were very similar. Based on independent validation of LDA classifications, we found that germination of Acacia and Corymbia seeds could be classified with over 85% accuracy, while it was about 80% for Banksia seeds . Regarding Acacia and Banksia seeds, we obtained similar classification accuracies of germinating and non-germinating seeds, but for Corymbia the classification accuracy associated with non-germinating seeds was considerably higher than for germinating seeds. We examined the relationships between days of ageing and classification accuracies, and it was revealed that the classification accuracy of : Corymbia seeds was above 80% at all time points, the classification accuracy of Acacia seeds was above 90% in the beginning and end of the study period but was below 65% around the time point with 50% germination, and Banksia seeds was below 80% during the time period with a marked decline in germination but around 90% in the beginning and end of the study period. In other words, seeds sampled during the gradual decrease in germination, in the transition from germination to non-germination, were generally classified with lower accuracy.Despite growing inTherest in use of reflectance based methods to determine the germination of horticulture and agriculture seeds, we are unaware of any published studies involving the assessment of native seeds for conservation or vegetation restoration purposes. We demonstrated that the germination of seeds can be accurately classified on the basis of commonly used classification methods, such as, LDA. Thus, it may be possible to replace time consuming and destructive germination tests with non-destructive reflectance based technologies as part of improved management of seed banks. The germination of Acacia and Banksia seeds decreased from above 90% to below 20% in about 10 days of experimental ageing. The decline in germination of Corymbia seeds was less pronounced from over 90% to about 10% in 20 days. Numerous studies have demonstrated successful use of reflectance-based spectroscopy as part of studies into seed germination. Shetty et al. used near-infrared spectroscopy to classify viable/ non-viable cabbage and radish seeds of different sizes with classification accuracies exceeding 90%. Similarly, Ahn et al.used a combination of hyperspectral imaging and fluorescence lighting to discriminate between viable and non-viable Brassica seeds with over 90%accuracy. Ahn et al. used the Fourier transform near-infrared reflectance technology technique to classify viable/non-viable water melon [Citrullus lanatus ] seeds and obtained classification accuracies also exceeding 90%. Finally, Esteve et al. used near-infrared reflectance technology to detect heat and frost damage to corn and to differentiate viable and non-viable corn and soybean seeds. The authors concluded that only heat damage could be accurately predicted. The exact associations between seed coat reflectance and primary seed metabolism are not known, so it is not possible to provide much more than speculations about the importance of certain changes in seed coat reflectance. However, some important insight may be gained by using knowledge gathered from reflectance studies of plant leaves. For instance, it is known that important plant pigments have maximum peaks at particular wavelengths: chlorophyll a , and carotenoids . In addition, there is a wealth of simple two spectral band indices used to estimate chlorophyll content in leaves, including : R430/ R680, R672/R550, R710/R760, R750/R550. Finally, there is a large body of research into use of reflectance based methods to quantify nitrogen content in leaves, and these have recently been reviewed. An important study analyzed the correlation between reflectance in spectral bands between R447–R1752 nm and leaf N accumulation in rice and wheat. The authors found that leaf N accumulation was strongly correlated with reflectance at R660, R810, and R870 nm.